Databricks is telling customers it has closed the gap between batch and streaming ML features: a claimed 200ms p99 latency from Kafka event to online feature store, via a new engine called Spark Real-Time Mode and a Lakebase write path built to avoid Postgres’s write-ahead-log amplification. That number matters because feature freshness has been the unglamorous bottleneck in production ML — fraud and personalization models are only as good as how recently their signals were updated, and until now vendors mostly solved this with bolted-on streaming infrastructure data scientists had to build themselves.
The awkward part is that this isn’t the only latency number in circulation. Snowflake, in its own blog on Snowflake Postgres, claims its online feature store beats “Databricks Online Feature Store, which is backed by Databricks Lakebase” by 2.5x on latency at equivalent throughput and 7x on QPS on comparable instances. Neither company published a shared benchmark methodology, instance sizing, or third-party audit, which is standard practice in this corner of the industry and standard reason for discounting both figures until an independent party reruns the test.
Databricks’ 200ms figure is a real architectural achievement — continuous-stage processing instead of microbatching, amortized checkpointing, a rewritten Postgres write path — but it’s also a number published unilaterally into a fight where Snowflake is publishing its own numbers straight back.
What’s more interesting than either vendor’s headline stat is the shared architectural bet: both companies are converging on Postgres-family databases (Lakebase, Snowflake Postgres) as the serving layer for low-latency ML, replacing the custom key-value stores that feature-store teams used to hand-roll. That consolidation is worth watching for anyone selling market or alternative data into ML pipelines — if warehouse vendors are commoditizing the serving tier, differentiation moves further upstream, to data freshness and lineage rather than infrastructure. Watch for independent benchmarks from customers running both stacks at production scale, and for whether either vendor discloses instance cost alongside the latency claims.
Databricks Feature Store brings real-time freshness to ML features: streaming aggregations from Kafka can now reach the online feature store with 200ms p99 latency, collapsing feature lag from minutes or hours to milliseconds.
